Logistics Simulation and Optimization
Predicting with Precision
Mathematik unterstützt die Logistik: Prognoseverfahren von Siemens sagen den Absatz von Produkten mit einer Genauigkeit bis 85 % voraus, und eine neue Software zur Routenplanung ist 1200-mal schneller als bisherige Verfahren.
Precise sales forecasts, produced with the help of such things as simulation programs, help optimize the value-creation chain for all types of products. The actual sales figures for Siemens mobile phones, including the effect of Christmas, are shown
Company sales representatives are only too aware that fear and insecurity can have a tremendous market impact. For instance, worries about the war against Iraq and SARS slowed cell phone sales, forcing companies to cut their sales forecasts for 2003. This, in turn, led to a drop in demand for flash memory chips, and many other items. So imagine how helpful it would be if up-to-date sales forecasts that instantly took account of such external factors could be created at the push of a button!
There isn't a business on earth that wouldn't embrace such a technology. Indeed, companies are having more and more trouble staying on top of supply chains that frequently stretch across many national borders. In order to optimally organize procurement, production, storage and sales with suppliers and customers, a company needs to know exactly how many cell phones, washing machines, televisions or automobiles it will sell in the coming weeks and months.
Better forecasts with neural networks
Neural networks that are based on the structure of the human brain are ideally suited for simulations that involve non-linear, complex situations. In general, artificial neurons are divided into three categories: input, hidden and output neurons. External data covering such things as economic factors flow through the input neurons. The hidden neurons, which can be arranged in several layers, process this data, and the output neurons provide the conclusion the level of future sales, for instance. Every neuron within a layer is connected to all of the neurons in the next layer by network parameters.
Neural networks learn from the data of the past. More specifically, the input neurons constantly receive sales updates: daily reports on total sales, prices, weather factors such as rain that send customers rushing into department stores, or seasonal factors such as Christmas. The neural network stores the business information in the network parameters. In the training mode, the learning algorithm changes the parameters until the network produces forecasts that deviate as little as possible from actual sales totals, and the input of additional information leads to no further improvement. At this point, the neural network is ready to make future sales forecasts.
Siemens uses "recurrent" networks in place of the widely used "feed-forward" networks to make sales forecasts. In feed-forward networks, the data flows in only one direction from the input layer to the output. In recurrent networks, on the other hand, signals from one layer are sent back to the one behind it. This makes the models more resistant to disruptive factors, and the network can be trained using less data.
In this context, cumputer-based simulation models could serve as an important decision-making tool. Conventional processes employ such mathematical methods of analysis as linear regression, progressive averages, and exponential smoothing. But a faster, better technology for simulating dynamic systems is neural networks. Neural networks have been used for a long time to produce sales, liquidity and stock forecasts. In this area, the Competence Center for Neural Computation at Siemens Corporate Technology (CT) is working with so-called recurrent neural networks (see box left). "Our systems can produce three-month sales forecasts that are 75 to 85 % accurate," says Dr. Ralph Neuneier, who is in charge of the "Learning Systems in Business Processes" section at CT. "Conventional time series analysis, on the other hand, is only 55 to 60 % accurate."
Many Factors. In an effort to improve sales forecasts, additional information about the value-creation chain is integrated into neural networks. The use of this information ensures that the forecast doesn't simply rely on such past data as previous sales totals. Instead, new information focuses on such factors as long-term supply contracts, use of production capacity, current inventory and the typical buying habits of major customers, who often place orders at the end of the quarter.
The market model is also given information about special seasonal events such as Christmas or planned marketing campaigns, along with economic indicators that measure such things as the business climate for sectors and companies. Neural networks can sort through all these factors and produce more accurate forecasts than competing systems.
Today, many CEOs and top managers are looking forward to switching from monthly to weekly planning cycles. "But this is still too long a period to react to sudden changes that could result in bottlenecks or delays and cause back-ups that stretch all the way to the end of the value-creation chain," says Dr. Rudolf Sollacher, who is in charge of self-organizing systems at Siemens CT. To eliminate this bull-whip effect, all of the partners in a process chain should be able to communicate directly and quickly with one another. In addition, associated software must be able to simulate not only the dynamics of the entire supply chain but also fluctuations in production and even the utilization level of individual machines.
Changes should be recorded quickly and automatically. This is the only way to take countermeasures that can prevent disruptions from turning into major problems. Siemens has already developed simulation software that can do the job.
Shorter, Faster Routes. Even when the right sales forecast has been put together, products still have to arrive at the right place and at the right time. In order to accomplish this, the route has to be optimally planned. The problem is that today's delivery vehicle-based navigation systems take the route step by step because their on-board computers have limited capacity a feature that can greatly restrict the search area depending on the dynamics of a situation.
A route plan with 500,000 sections from Moscow to the Canary Islands can be produced in under a thousandth of a second
For example, such systems may first direct a driver from a downtown area to a beltway, then to a major artery, and finally to an interstate highway. The computer only begins to refine the search once the driver approaches his or her destination. "This method rarely shows the driver the shortest or quickest route," says Professor Ulrich Lauther, who is in charge of Efficient Algorithms in Networks at Siemens CT.
With these problems in mind, Lauther's team has developed a new process that is capable of computing the optimal route from the Canary Islands to Moscow within a milli-second on a notebook that's all the time it takes for a new route planner to handle the 500,000 sections of the trip, including metropolitan areas, roads, bridges and ferry connections (see graphic above).
In the process, virtual signposts are set up at every fork in the road, pointing out the highways that will produce the shortest and fastest trip. It takes about 20 minutes to set up the signs. This process must be performed at the start (or whenever the digital map changes because of things like blocked roads or traffic jams). Afterward, though, the route calculation system works 1,200 times faster than conventional methods do.
Admittedly because the 20-minute set-up period is impractical the software is not really equipped to deal with the requirements of vehicle navigation systems.
However, the software is ideal for a system based on a central computer that can regularly integrate the latest traffic reports into new calculations. A server of this sort could, for instance, be used in a logistics company or a trucking operation, or it could be used on the Internet for route planning purposes. Such cost factors as tolls or time windows for deliveries to customers can also be taken into consideration.
Preferred Roads. One problem with the use of digital maps is what actually constitutes a preferred road. "These are not always expressways or federal highways, because situations arise that lead to long detours when these particular routes are used," Lauther says.
That's why Siemens is working on software that will find the important roads for optimal route planning independent of the information provided by map makers. If one highway regularly appears in various route calculations, it is particularly important and is highlighted on the digital map. It thus gains priority in the route planning that follows. As a result, the reliability of the road classification rises compared with the information provided by the map maker.
A licensee is already using Siemens' route planner to optimize the trips of a logistics company.
But the route planner is not just suited for highways. It also can be used in communications networks. After all, there are many routes that a message can take. A communications network can therefore be considered to be a map and the various network junctions can be seen as intersections. Thus, a defective or heavily used cable can be integrated like a blocked street into route precalculations in order to produce the optimal cable connection.
Sylvia Trage